Fuzzy Regions for Handling Uncertainty in Remote Sensing Image Segmentation

1
Fuzzy Regions for Handling Uncertainty in Remote
Sensing Image Segmentation
Ivan Lizarazo, (a) and Paul Elsner (b) (a)
Department of Cadastral Engineering
University Distrital, Bogota, Colombia (b) School
of Geography, Birkbeck College, University
of London, UK
2
Agenda
1. Introduction2. Case Study Urban
land-cover classification 3. Results4.
Conclusions
3
Introduction

• Geographic Object-based Image Analysis
• - alternative to pixel-wise classification.
• - includes contextual and geometric
  information.
• - key steps
• (1) group pixels into segments.
• (2) evaluate segments properties.

Fuzzy Regions
4
Introduction (2)

• Geographic Object-based Image Analysis

Pre-processed pixels
Segmentation
Image Objects
Attributes Assessment
Attributes Vector
Classification
Ground Objects
Fuzzy Regions
5
Introduction (3)

• Discrete Image Segmentation
• Image is subdivided into discrete objects with
  well defined boundaries

Fuzzy Regions
6
Introduction (4)

• Problems of Discrete Image Segmentation
• Noisy images and pixel mixed may produce
• meaningless image-objects.
• Geographic objects are not always
• discrete features.
• Establishing a correspondence between
• image-objects and real-world objects
• is a time-consuming process.

I. Lizarazo
7
Introduction (5)

• Continuous Image Segmentation
• Image is subdivided into fuzzy objects
• with degrees of membership to classes

A
B
Input image
C
Segmented image
I. Lizarazo
8
Case Study Classification of urban land-cover

• Geographic Area Washington DC-Mall
• Data HYDICE Imagery
• 191 spectral bands
• 3 meters spatial resolution
• 1280 x 307 pixels
• Ground Reference
• Training dataset 704 pixels
• Testing dataset 1193 pixels

http//cobweb.ecn.purdue.edu/landgreb/Hyperspectr
al.Ex.html
Fuzzy Regions
9
Hydice Imagery
I. Lizarazo
10
Methods

Fuzzy Regions
11
Methods

Fuzzy Regions
12
Methods Segmentation

• Support Vector Machine (SVM)
• Given training data (xi, yi) find
• a function f(x) that has at most
• e deviation from the targets yi
• Transformation of the original space into a
  higher dimension using a kernel function k(x,xi)

Fuzzy Regions
13
Methods Segmentation

• SVM Kernel Radial Basis Function
• Automated SVM parameterization

Implementation libsvm (R package)
I. Lizarazo
14
Methods Attribute Assessment

• - Overlapping Index (Lambert and Grecu, 2003)
• Confusion Index (Burrough et al, 1997)

Fuzzy Regions
15
Methods Defuzzification

Fuzzy Regions
Fuzzy Regions Intensified
CL-2
Fuzzy Union Operation
SVM-based Classification
CL-3
CL-1
Land-cover Classes
I. Lizarazo
16
Methods Defuzzification

Fuzzy Regions
Fuzzy Union Operation
CL-1
Land-cover Classes
I. Lizarazo
17
Methods Defuzzification

Fuzzy Regions
Fuzzy Regions Intensified
SVM-based Classification
CL-3
Land-cover Classes
I. Lizarazo
18
Methods Defuzzification

• CL2 - CL3
• SVM-based classification There is a separating
  hyperplane which maximises the margin between
  classes

I. Lizarazo
19
Results Fuzzy Image-Regions

Road
Roof
Shadow
OI
I. Lizarazo
20
Results Fuzzy Image-Regions

Grass
Trees
Water
Trail
I. Lizarazo
21
Results Land-cover classification

CI
CL-1
CL-2
Reference
Fuzzy Regions
22
Results Classification Accuracy

• CL-2
• Percentage of Correct Classification 87

Fuzzy Regions
23
Conclusions

• Fuzzy Image Segmentation alternative for
  handling ambiguous information
• Automated SVM parameterisation may help users to
  produce accurate classifications
• R provides useful functionalities for remote
  sensing image analysis

Questions?
I. Lizarazo